MusGU+ Evaluation: JAM

The documented training setup requires high-end GPU resources. All training stages were carried out using 8 H100 GPUs, and no practical CPU-only or consumer-GPU adaptation pathway is demonstrated.

No clear information about the dataset size is provided. The model was trained on a dataset of ~54k hours of audio, which makes adaptation infeasible with musician’s own data.

JAM provides code and documentation for pretraining, supervised fine-tuning, and direct preference optimization, as well as pretrained model checkpoints. However, the released training workflows assume preprocessed representations and metadata and do not provide a complete end-to-end preprocessing pathway from a musician’s own audio data.

Adaptation requires technical expertise. All training and preprocessing rely on command-line implementation. No user-friendly adaptation interface is provided.

Redistribution of an adapted model based on JAM is subject to a dual license structure combining Project Jamify License and Stability AI Community Licence. Use, modification and redistribution are strictly restricted to non-commercial, academic, and entertainment purposes and require explicit attribution.

There is a hosted HuggingFace web demo, but it is currently unavailable. However, the Gradio code is accessible and can be run locally, though it requires installation and environment setup by the user.

JAM’s source code is openly available and can be freely used without limits, paywalls, or subscriptions.

No clear information about the time required for generation is provided. No reference to real-time generation is provided in the documentation or the research article.

There are no clear instructions or specific tools to integrate JAM to a musician’s workflow.

Outputs generated with JAM are restricted to non-commercial use and require explicit attribution acknowledging Stability AI licensing. Users are responsible for ensuring generated outputs do not infringe on copyrighted material.

Support exists via GitHub issues and documentation, but there is no dedicated musician-oriented support space.

JAM can be conditioned by lyrics (including word- and phoneme-level timing) and style (via text or audio).

JAM supports precise time-varying control over vocal structure, with explicit word- and phoneme-level timing and duration inputs that directly guide generation at the latent space. However, no time-varying control is provided for the accompaniment or instrumental structure.

JAM separates lyrics timing, global duration, and style conditioning through explicit control pathways, enabling partial disentanglement of vocal structure. However, musical attributes are not independently controllable, and style remains an entangled representation, limiting fully interpretable, attribute-specific control.

JAM provides several generation control parameters, including global duration, separate guidance strengths for lyrics and style, and diffusion step configuration, allowing users to directly influence generation behavior. However, direct manipulation of internal representations is not supported.